Using scientifically and statistically sufficient statistics in comparing image segmentations.

Area of Science:

• Medical imaging
• Radiotherapy
• Statistical modeling

Background:

• Automatic computer segmentation in 3D aids radiotherapy treatment planning.
• High-dimensional, low-sample-size (HDLSS) data from human vs. computer segmentation accuracy poses statistical challenges.
• Overcoming the curse of dimensionality is crucial for reliable inference.

Purpose of the Study:

• To develop a statistical strategy for analyzing HDLSS data in medical image segmentation.
• To compare the accuracy and consistency of human observers versus computer segmentation of kidneys in CT scans.
• To assess the clinical significance of segmentation discrepancies.

Main Methods:

• Reduced 3D distance data to histograms for individual modeling.
• Employed non-parametric kernel density estimation to analyze distributional patterns.
• Utilized Gaussian distribution modeling on transformed distance data.
• Represented each histogram with estimated distribution parameters to mitigate HDLSS issues.
• Applied classical statistical methods to compare observer and computer agreement.

Main Results:

• The statistical approach successfully reduced complex data to manageable parameters.
• Human observers showed significantly less disagreement with each other than with computer segmentation.
• All observed disagreements were clinically unimportant relative to kidney size.
• The hierarchical modeling approach yielded statistically and scientifically sufficient variables.

Conclusions:

• The developed statistical strategy effectively addresses HDLSS data challenges in medical imaging.
• Computer segmentation shows potential but human observers remain more consistent.
• The methodology is applicable to other high-throughput technologies in various scientific fields.